1# RECORD 4c31e259cf5625db8bd0cbbcae78c68fed2c81642# TITLE: Text-augmented house valuation: using embedding vectors and SHAP values3# YEAR: 2026.0 VENUE: Aestimum DOI: 10.36253/aestim-198244# AUTHORS: Lee Changro5# PAGES: 14 CANDIDATE TABLE PAGES (1-based): [6, 8, 14]6# ABSTRACT: With the recent advent of large language models, text data can now be readily processed and analyzed across various sectors; traditionally, this was not the case, and text data had proven challenging to analyze. The housing market uses and features a vast amount of textual data, such as property descriptions and neighborhood trend reports, but has yet to fully exploit this potential. This study has two objectives: first, to enhance house valuation accuracy by integrating text information into valuation models; and second, to interpret how text information contributes to improved performance. We convert the environmental descriptions of each house into embedding vectors, which are then incorporated into regression and extreme gradient boosting (XGB) models to estimate house prices. Additionally, we interpret the XGB model’s results using Shapley Additive explanations (SHAP) values. The embedding vectors significantly improved the performances of both the regression and XGB models, and the embedding vector values aligned well with the semantic meaning of the environmental descriptions of each house. This study contributes to the literature by deepening our understanding of embedding vectors in the context of house valuations and housing markets.789===== PAGE 1 (score 4, numeric-paren 0, table-words 0, var-words 4) =====1011 12 131 14AESTIMUM 15 16JUST ACCEPTED MANUSCRIPT 17 18 19 20Text-augmented house valuation: using embedding vectors and SHAP values 21 22Lee Changro 23 24Department of Real Estate, Kangwon National University, South Korea 25Email: spatialstat@naver.com 26 27 28 29 30This article has been accepted for publication and undergone full peer review but has not been through 31the copyediting, typesetting, pagination and proofreading process, which may lead to differences 32between this version and the Version of Record. 33 34 35 36 37 38 39 40 41 42 43Please cite this article as: 44Changro, L. (2026). Text-augmented house valuation: using embedding vectors and SHAP values. 45Aestimum, Just Accepted. 46DOI: 10.36253/aestim-19824 47JUST ACCEPTED 484950===== PAGE 6 (score 31, numeric-paren 15, table-words 2, var-words 11) =====5152 53 546 55AESTIMUM 56 57JUST ACCEPTED MANUSCRIPT 58 59 60Figure 2. Flow of research. 61 62To convert the environmental descriptions into embedding vectors, three embedding models are employed: a word-63embedding model (glove-wiki-gigaword-50) and two sentence-embedding models (paraphrase-MiniLM-L12-v2 and 64paraphrase-mpnet-base-v2). The glove-wiki-gigaword-50 is pre-trained and represents words as 50-dimensional vectors 65using data from Wikipedia and the Gigaword corpus. Paraphrase-MiniLM-L12-v2 and paraphrase-mpnet-base-v2, are 66referred to as sentence embeddings A and B, respectively; both are pretrained models from the Sentence Transformer 67library. Sentence embedding A is a lightweight transformer model that produces 384-dimensional embedding vectors, 68whereas sentence embedding B, which is an enhanced version of BERT and RoBERTa, generates 768-dimensional 69embedding vectors4. 70Evaluating and selecting the relevant embedding models is challenging (Torregrossa et al., 2021). In this study, model 71selection was guided by two complementary criteria: semantic similarity and downstream valuation performance. To 72illustrate the relative representational quality of the three embedding models, the most common environmental description 73in the dataset, “Remote and rural area” (455 houses, 27% of the total), was used as a reference, and the five most similar 74descriptions were identified using cosine distance and presented in Table 3. While relying on a single reference sentence 75may not fully capture the diversity of environmental descriptions across the dataset, this comparison was intended as an 76interpretable demonstration rather than the sole basis for selection. Descriptions in word embedding showed less similarity 77to the reference, whereas those in sentence embeddings A and B were more similar. Ultimately, sentence embedding B 78was chosen for use in the subsequent valuation models as it outperformed sentence embedding A in a downstream model 79performance (i.e., XGB). 80 81Table 3. Top five environmental descriptions similar to “Remote and rural area”. 82No. 83Word embedding 84Sentence embedding A 85Sentence embedding B 861 87Pure farming area 88(0.321) 89Rural area near a village 90(0.074) 91Rural area 92(0.062) 932 94Existing industrial area 95(0.362) 96Rural area 97(0.083) 98Rural area on a mountainside 99(0.149) 1003 101High-density commercial district 102(0.415) 103Rural area on a mountainside 104(0.167) 105Pure farmland 106(0.212) 1074 108Existing residential area 109(0.428) 110Rural area nearby the coast 111(0.186) 112Rural area near the suburbs 113(0.268) 1145 115Emerging residential area 116(0.429) 117Rural area near the suburbs 118(0.213) 119Rural area nearby the coast 120(0.281) 121Note: Numerical values indicate the cosine distance (= 1 – cosine similarity), where a value closer to zero indicates greater similarity. 122 1234 Details on these models are available at https://www.sbert.net/docs/sentence_transformer/pretrained_models.html. Prior to 124embedding, the environmental descriptions underwent text preprocessing steps. These included lowercasing all text, removing 125punctuation and special characters, and stripping leading and trailing whitespace. Tokenization was handled internally by each 126embedding model: the word embedding model tokenizes text at the word level, splitting descriptions into individual word tokens, while 127the two sentence embedding models employ sub-word tokenization via their respective built-in tokenizers from the Sentence 128Transformers library, which handle out-of-vocabulary terms more robustly. 129130131===== PAGE 7 (score 12, numeric-paren 0, table-words 0, var-words 12) =====132133 134 1357 136AESTIMUM 137 138JUST ACCEPTED MANUSCRIPT 139 140 141Sentence embedding B consists of 768-dimensional vectors which can be inefficient if used in their entirety in 142subsequent valuation models. To enhance efficiency, sentence embedding B is reduced to two dimensions using principal 143component analysis (PCA). The explained variance ratio for these two dimensions (two components) is 37%, with the 144first and second components accounting for 22% and 15%, respectively: hereafter referred to as embedding components 1451 and 25. 1463.3 Valuation models 147Three valuation models are used to estimate house prices: OLS without text information, OLS with text information, 148and XGB with text information. The first model acts as the baseline. XGB is a widely used machine learning model that 149has demonstrated excellent performance in various studies (Alshboul et al., 2022; Chen et al., 2015; Osman et al., 2021). 150An additional benefit of using XGB is the ease of computing Shapley Additive explanations (SHAP) values, which can 151be efficiently calculated for tree-based models, including gradient boosting. 152The SHAP value begins by examining all possible subsets of explanatory variables, commonly known as coalitions. 153For each explanatory variable, its marginal contribution to the prediction is calculated when added to each possible 154coalition of the other explanatory variables. The SHAP value for an explanatory variable is the average of its marginal 155contributions across all possible coalitions (Lundberg and Lee, 2017). A positive SHAP value indicates that the 156explanatory variable positively contributes to the prediction or pushes it toward a higher value, whereas a negative SHAP 157value indicates the opposite effect. The absolute value of SHAP value indicates the strength of the contribution of the 158explanatory variable. A SHAP value of zero indicates that the explanatory variable has no impact on the prediction6. 159To estimate house prices, Equation (1) was used: 160 161𝑃𝑟𝑖𝑐𝑒 (𝑙𝑜𝑔−𝑡𝑟𝑎𝑛𝑠𝑓𝑜𝑟𝑚𝑒𝑑) = ℎ𝑜𝑢𝑠𝑒 𝑡𝑦𝑝𝑒+ 𝑠𝑖𝑡𝑒 𝑎𝑟𝑒𝑎+ 𝑓𝑙𝑜𝑜𝑟 𝑎𝑟𝑒𝑎+ 𝑠𝑡𝑜𝑟𝑦+ 𝑟𝑜𝑎𝑑+ 𝑠𝑖𝑡𝑒 𝑠ℎ𝑎𝑝𝑒+162+ 𝑏𝑢𝑖𝑙𝑑𝑖𝑛𝑔 𝑓𝑟𝑎𝑚𝑒+ 𝑝𝑟𝑜𝑝𝑒𝑟𝑡𝑦 𝑎𝑔𝑒+ ℎ𝑜𝑢𝑠𝑒 𝑏𝑒𝑎𝑟𝑖𝑛𝑔+ 𝑒𝑚𝑏𝑒𝑑𝑑𝑖𝑛𝑔 𝑐𝑜𝑚𝑝𝑜𝑛𝑒𝑛𝑡 1 +163 + 𝑒𝑚𝑏𝑒𝑑𝑑𝑖𝑛𝑔 𝑐𝑜𝑚𝑝𝑜𝑛𝑒𝑛𝑡 2 164 165 166 167 168 169 170 171 172 173 (1) 174 175From the explanatory variables available in the benchmark house data, 11 were selected after reviewing their 176coefficients and significance. House type indicates the subcategory of a house, such as detached, attached, rowhouse, 177mixed type, and other variations. Site area refers to the area of the house site, whereas floor area is the total floor space 178within the house. Story denotes the number of floors; road categorizes the size of the road, including main, middle, narrow 179roads, and other variations; and site shape describes whether the house site is regular or irregular. Building frame includes 180types like reinforced concrete (RC), wooden frames, steel frames, brick masonry, and others. Property age is the number 181of years since a house was constructed and house bearing indicates the direction that the house faces (south or non-south). 182The final two variables were the principal components derived from the embedding vectors. 183The XGB implementation details are as follows. The maximum depth of the decision tree was set to 3. Increasing this 184value can result in a more complex model and potentially cause overfitting. A learning rate of 1.0 was used. The number 185of boosting rounds was set to 20. In addition, the squared error was selected as the loss function7. 186 187 188 1894. Results and discussion 190 1915 The number of PCA components was selected based on an evaluation of a downstream model performance using XGB. Models 192incorporating 3, 5, and 7 components were compared against a two-component baseline using mean squared error (MSE). No 193meaningful improvement was observed beyond two components. Additionally, beyond the first two components, each additional 194component individually explained less than 10% of the variance, with diminishing marginal returns. Incorporating more components 195also increase the risk of overfitting, particularly given the relatively small sample size (763 in Seongdong-gu and 907 in Gangjin-gun). 196Thus, a two-component representation was retained as the reasonable specification. 1976 Details on the SHAP value implementation are provided in Lundberg and Lee (2017). 1987 These hyperparameters were determined through evaluation on a validation dataset, where key parameters (tree depth, learning rate, 199and number of boosting rounds) were varied and the resulting goodness of fit was assessed. The reported values reflect the configuration 200that yielded the best performance. We acknowledge that a more systematic tuning approach, such as grid search, was not employed, 201which remains a limitation of the current study. 202203204===== PAGE 8 (score 34, numeric-paren 17, table-words 5, var-words 4) =====205206 207 2088 209AESTIMUM 210 211JUST ACCEPTED MANUSCRIPT 212 2134.1 Text-driven model improvement 214Table 4 presents the results of the OLS analysis. In both Seongdong-gu and Gangjin-gun, embedding components 1 215and 2 significantly enhanced the model performance. The adjusted R² values increased from 0.508 to 0.558 and from 2160.633 to 0.751 in both regions, respectively. This indicates that text information, specifically environmental descriptions, 217contributes to improving valuation accuracy. 218 219Table 4. Results of the OLS analysis. 220Variables 221Seongdong-gu 222Gangjin-gun 223Descriptions 224not included 225Descriptions 226included 227Descriptions 228not included 229Descriptions 230included 231Intercept 2324.791* 2334.581* 2342.394* 2352.224* 236House type: rowhouse 237- 238- 2390.381+ 2400.455+ 241House type: attached 242-0.082+ 243-0.010+ 244- 245- 246House type: detached 2470.010+ 2480.018+ 2490.306+ 2500.264+ 251House type: mixed type 252-0.468* 253-0.553* 254-0.414+ 255-0.623* 256House type: other 257-0.706* 258-0.611* 2590.060+ 2600.208+ 261Site area 2620.009* 2630.009* 2640.000* 2650.000* 266(Skipped for brevity) 267… 268… 269… 270… 271Embedding component 1 272- 273-0.265* 274- 275-0.393* 276Embedding component 2 277- 2780.592* 279- 280-0.090* 281Adjusted R2 2820.508+ 2830.558+ 2840.633+ 2850.751+ 286* Significant at the 5.0% level. The complete table is provided in the Appendix (Table A1). 287 288Figure 3 shows the goodness-of-fit of the test dataset for the three models: OLS without text information, OLS with 289text information, and XGB with text information. In both regions, the OLS model incorporating environmental 290descriptions demonstrated a better fit than the OLS model without such information. Furthermore, the XGB model 291exhibited the best fit among the three models in both regions, making it a reliable choice for more detailed interpretations8. 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 3158 In Seongdong-gu, the difference between OLS without text and with text may not be obvious, but the improvement from OLS with 316text to XGB with text is discernible: the latter model’s predictions align more closely with observed values. In Gangjin-gun, the 317increasing alignment between predictions and observed values (from OLS without text, to OLS with text, and finally to XGB with text) 318is noticeable. 319320321===== PAGE 9 (score 4, numeric-paren 0, table-words 0, var-words 4) =====322323 324 3259 326AESTIMUM 327 328JUST ACCEPTED MANUSCRIPT 329 330Figure 3. Goodness-of-fit. 3314.2 SHAP value interpretation 332The XGB model demonstrated the best fit, as shown in Figure 3. Figure 4 presents the importance of the explanatory 333variables, as determined by the SHAP values of the XGB model. Seongdong-gu, a highly urbanized area with scarce land, 334identifies site area as the most critical factor in estimating house prices. In contrast, Gangjin-gun, a rural area with 335abundant land, finds building-related factors such as property age to be more significant in house price estimations. In 336both regions, embedding components, especially the first one (embedding component 1), rank third in Seongdong-gu and 337second in Gangjin-gun9. 338 3399 As shown in Figure 4, embedding component 2 ranked sixth in both Seongdong-gu and Gangjin-gun, indicating a non-negligible role 340in house valuation. As noted earlier, the two embedding components together capture 37% of the variance in the sentence embeddings, 341with components 1 and 2 accounting for 22% and 15%, respectively. By construction, PCA produces orthogonal components, meaning 342embedding component 2 captures a dimension of linguistic variation in environmental descriptions that is not represented by component 3431. The semantic interpretation of component 2 is less straightforward than that of component 1, and therefore we focus on the 344interpretation of component 1 in this study. 345346347===== PAGE 14 (score 74, numeric-paren 56, table-words 2, var-words 6) =====348349 350 35114 352AESTIMUM 353 354JUST ACCEPTED MANUSCRIPT 355 356Appendix 357Table A1. Results of the OLS analysis. 358Variables 359Seongdong-gu 360Gangjin-gun 361Descriptions 362not included 363Descriptions 364included 365Descriptions 366not included 367Descriptions 368included 369Intercept 3704.791* 3714.581* 3722.394* 3732.224* 374House type: rowhouse 375- 376- 3770.381+ 3780.455+ 379House type: attached 380-0.082+ 381-0.010+ 382- 383- 384House type: detached 3850.010+ 3860.018+ 3870.306+ 3880.264+ 389House type: mixed type 390-0.468* 391-0.553* 392-0.414+ 393-0.623* 394House type: other 395-0.706* 396-0.611* 3970.060+ 3980.208+ 399Site area 4000.009* 4010.009* 4020.000* 4030.000* 404Floor area 405-0.001* 406-0.002* 4070.001* 4080.001* 409Story 4100.084* 4110.087* 4120.319* 4130.118+ 414Road: middle 4150.204* 4160.204* 4170.334+ 4180.440+ 419Road: narrow 4200.184* 4210.226* 422-0.295+ 423-0.022+ 424Road: narrower 4250.288* 4260.314* 427-0.696+ 428-0.206+ 429Road: narrowest 4300.120+ 4310.177* 432-0.654+ 433-0.195+ 434Site shape: regular 4350.144* 4360.136* 4370.204* 4380.108* 439Building frame: brick 4400.246* 4410.284* 4420.792* 4430.734* 444Building frame: RC 4450.375* 4460.438* 4471.475* 4481.385* 449Building frame: low quality wood 4500.197+ 4510.239+ 4520.153* 4530.237* 454Building frame: high quality wood 455- 456- 4571.730* 4581.786* 459Building frame: panel 460- 461- 4620.740* 4630.862* 464Building frame: steel 465- 466- 4671.142* 4681.304* 469Property age 470-0.005* 471-0.005* 472-0.010* 473-0.011* 474House bearing: south 475-0.063* 476-0.038+ 4770.021+ 4780.051+ 479Embedding component 1 480- 481-0.265* 482- 483-0.393* 484Embedding component 2 485- 4860.592* 487- 488-0.090* 489Adjusted R2 4900.508+ 4910.558+ 4920.633+ 4930.751+ 494*: Significant at the 5.0% level. 495